Separable Ky-Fan spectral regularization / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "tabular",
4 "model": "mlp_tiny",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.006,
10 "wd": 0.0001
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "wd": 0.0
17 },
18 "mean": 69.17356395721436
19 },
20 {
21 "cfg": {
22 "lr": 0.001,
23 "wd": 0.0001
24 },
25 "mean": 69.16811275482178
26 },
27 {
28 "cfg": {
29 "lr": 0.003,
30 "wd": 0.0
31 },
32 "mean": 15.44080638885498
33 },
34 {
35 "cfg": {
36 "lr": 0.003,
37 "wd": 0.0001
38 },
39 "mean": 15.364070892333984
40 },
41 {
42 "cfg": {
43 "lr": 0.006,
44 "wd": 0.0
45 },
46 "mean": 12.937470436096191
47 },
48 {
49 "cfg": {
50 "lr": 0.006,
51 "wd": 0.0001
52 },
53 "mean": 12.618298768997192
54 }
55 ],
56 "full": {
57 "mean": 12.842816472053528,
58 "std": 1.3793130820674717,
59 "per_seed": [
60 13.872417449951172,
61 12.713014602661133,
62 10.938212394714355,
63 12.94955062866211,
64 10.740423202514648,
65 12.51738452911377,
66 14.032771110534668,
67 14.978757858276367
68 ],
69 "n": 8
70 },
71 "idea_union_sweep": [
72 {
73 "cfg": {
74 "lr": 0.001,
75 "wd": 0.0,
76 "lam": 0.001,
77 "k": 1
78 },
79 "mean": 198.39544296264648
80 },
81 {
82 "cfg": {
83 "lr": 0.001,
84 "wd": 0.0,
85 "lam": 0.003,
86 "k": 1
87 },
88 "mean": 198.3975486755371
89 },
90 {
91 "cfg": {
92 "lr": 0.001,
93 "wd": 0.0,
94 "lam": 0.001,
95 "k": 4
96 },
97 "mean": 198.3946418762207
98 },
99 {
100 "cfg": {
101 "lr": 0.001,
102 "wd": 0.0001,
103 "lam": 0.001,
104 "k": 1
105 },
106 "mean": 198.27414321899414
107 },
108 {
109 "cfg": {
110 "lr": 0.001,
111 "wd": 0.0001,
112 "lam": 0.003,
113 "k": 1
114 },
115 "mean": 198.27565002441406
116 },
117 {
118 "cfg": {
119 "lr": 0.001,
120 "wd": 0.0001,
121 "lam": 0.001,
122 "k": 4
123 },
124 "mean": 198.2729377746582
125 },
126 {
127 "cfg": {
128 "lr": 0.003,
129 "wd": 0.0,
130 "lam": 0.001,
131 "k": 1
132 },
133 "mean": 20.76108717918396
134 },
135 {
136 "cfg": {
137 "lr": 0.003,
138 "wd": 0.0,
139 "lam": 0.003,
140 "k": 1
141 },
142 "mean": 20.76112723350525
143 },
144 {
145 "cfg": {
146 "lr": 0.003,
147 "wd": 0.0,
148 "lam": 0.001,
149 "k": 4
150 },
151 "mean": 20.760785341262817
152 },
153 {
154 "cfg": {
155 "lr": 0.003,
156 "wd": 0.0001,
157 "lam": 0.001,
158 "k": 1
159 },
160 "mean": 20.707895040512085
161 },
162 {
163 "cfg": {
164 "lr": 0.003,
165 "wd": 0.0001,
166 "lam": 0.003,
167 "k": 1
168 },
169 "mean": 20.70781660079956
170 },
171 {
172 "cfg": {
173 "lr": 0.003,
174 "wd": 0.0001,
175 "lam": 0.001,
176 "k": 4
177 },
178 "mean": 20.707414388656616
179 },
180 {
181 "cfg": {
182 "lr": 0.006,
183 "wd": 0.0,
184 "lam": 0.001,
185 "k": 1
186 },
187 "mean": 15.14272427558899
188 },
189 {
190 "cfg": {
191 "lr": 0.006,
192 "wd": 0.0,
193 "lam": 0.003,
194 "k": 1
195 },
196 "mean": 15.143461465835571
197 },
198 {
199 "cfg": {
200 "lr": 0.006,
201 "wd": 0.0,
202 "lam": 0.001,
203 "k": 4
204 },
205 "mean": 15.142550706863403
206 },
207 {
208 "cfg": {
209 "lr": 0.006,
210 "wd": 0.0001,
211 "lam": 0.001,
212 "k": 1
213 },
214 "mean": 14.998723030090332
215 },
216 {
217 "cfg": {
218 "lr": 0.006,
219 "wd": 0.0001,
220 "lam": 0.003,
221 "k": 1
222 },
223 "mean": 15.001027822494507
224 },
225 {
226 "cfg": {
227 "lr": 0.006,
228 "wd": 0.0001,
229 "lam": 0.001,
230 "k": 4
231 },
232 "mean": 14.997964859008789
233 }
234 ]
235 },
236 "idea": {
237 "mean": 15.766902923583984,
238 "std": 1.2326099075628074,
239 "per_seed": [
240 14.694108963012695,
241 15.323976516723633,
242 15.881936073303223,
243 14.091837882995605,
244 17.229509353637695,
245 14.569803237915039,
246 16.688566207885742,
247 17.655485153198242
248 ],
249 "n": 8
250 },
251 "comparison": {
252 "delta_mean": 2.9240864515304565,
253 "idea_wins": 0,
254 "n_pairs": 8,
255 "per_seed_diffs": [
256 0.8216915130615234,
257 2.6109619140625,
258 4.943723678588867,
259 1.142287254333496,
260 6.489086151123047,
261 2.0524187088012695,
262 2.655795097351074,
263 2.676727294921875
264 ],
265 "p_value": 0.0081,
266 "mde": 1.591959503517395,
267 "mde_rel_pct": 12.395719482416972,
268 "verdict": "idea worse (significant)",
269 "system_worked": false
270 },
271 "verification": {
272 "trials": 40,
273 "failures": 0,
274 "worst_cumulative_gap": 2.7284841053187847e-12,
275 "max_eigenvalue_ratio_mean": 0.7408118250286352
276 },
277 "idea_best_cfg": {
278 "lr": 0.006,
279 "wd": 0.0001,
280 "lam": 0.001,
281 "k": 4
282 },
283 "mechanism_signature": {
284 "quantity": "exact operator norm / separable R1 bound on trained models",
285 "predicted": "ratio <= 1 by the Ky-Fan upper bound",
286 "observed_mean_ratio": 0.5401321259511417,
287 "observed": [
288 {
289 "bound_R4": 0.9672489702701569,
290 "bound_R1": 3.6167073249816895,
291 "exact_norm": 2.102071504543225,
292 "ratio_R1": 0.5812113935854257
293 },
294 {
295 "bound_R4": 1.1515326688687006,
296 "bound_R1": 4.763582706451416,
297 "exact_norm": 2.377279565483332,
298 "ratio_R1": 0.4990528583168576
299 }
300 ],
301 "confirmed": true
302 },
303 "stage2_note": "Baseline and idea use same tabular data, epochs, batch size, Adam, learning-rate/weight-decay union; idea changes only the first hidden operator and adds Ky-Fan penalty."
304}